Who Should Stop Unethical A.I.?
At artificial-intelligence conferences, researchers are increasingly alarmed by what they see.
By Matthew Hutson
February 15, 2021
In computer science, the main outlets for peer-reviewed research are not journals but conferences, where accepted papers are presented in the form of talks or posters. In June, 2019, at a large artificial-intelligence conference in Long Beach, California, called Computer Vision and Pattern Recognition, I stopped to look at a poster for a project called Speech2Face. Using machine learning, researchers had developed an algorithm that generated images of faces from recordings of speech. A neat idea, I thought, but one with unimpressive results: at best, the faces matched the speakers’ sex, age, and ethnicity—attributes that a casual listener might guess. That December, I saw a similar poster at another large A.I. conference, Neural Information Processing Systems (Neurips), in Vancouver, Canada. I didn’t pay it much mind, either.
Not long after, though, the research blew up on Twitter. “What is this hot garbage, #NeurIPS2019?” Alex Hanna, a trans woman and sociologist at Google who studies A.I. ethics, tweeted. “Computer scientists and machine learning people, please stop this awful transphobic shit.” Hanna objected to the way the research sought to tie identity to biology; a sprawling debate ensued. Some tweeters suggested that there could be useful applications for the software, such as helping to identify criminals. Others argued, incorrectly, that a voice revealed nothing about its speaker’s appearance. Some made jokes (“One fact that this should never have been approved: Rick Astley. There’s no way in hell that their [system] would have predicted his voice out of that head at the time”) or questioned whether the term “transphobic” was a fair characterization of the research. A number of people said that they were unsure of what exactly was wrong with the work. As Hanna argued that voice-to-face prediction was a line of research that “shouldn’t exist,” others asked whether science could or should be stopped. “It would be disappointing if we couldn’t investigate correlations—if done ethically,” one researcher wrote. “Difficult, yes. Impossible, why?”
Some of the conversation touched on the reviewing and publishing process in computer science. “Curious if there have been discussions around having ethics review boards at either conferences or with funding agencies (like IRB) to guide AI research,” one person wrote. (An organization’s institutional review board, or I.R.B., performs an ethics review of proposed scientific research.) Many commenters pointed out that the stakes in A.I. research aren’t purely academic. “When a company markets this to police do they tell them that it can be totally off?” a researcher asked. I wrote to Subbarao Kambhampati, a computer scientist at Arizona State University and a past president of the Association for the Advancement of Artificial Intelligence, to find out what he thought of the debate. “When the ‘top tier’ AI conferences accept these types of studies,” he wrote back, “we have much less credibility in pushing back against nonsensical deployed applications such as ‘evaluating interview candidates from their facial features using AI technology’ or ‘recognizing terrorists, etc., from their mug shots’—both actual applications being peddled by commercial enterprises.” Michael Kearns, a computer scientist at the University of Pennsylvania and a co-author of “The Ethical Algorithm,” told me that we are in “a little bit of a Manhattan Project moment” for A.I. and machine learning. “The academic research in the field has been deployed at massive scale on society,” he said. “With that comes this higher responsibility.”